]> git.sesse.net Git - ffmpeg/blobdiff - libavcodec/vaapi_encode.c
vaapi_encode: Add MPEG-2 support
[ffmpeg] / libavcodec / vaapi_encode.c
index ddd03084921875e36e0da51c31aebdf0b783c69d..949c29846c621f312a9f5fcb9169ec7e5c1b691e 100644 (file)
@@ -109,10 +109,10 @@ static int vaapi_encode_wait(AVCodecContext *avctx,
     }
 
     av_log(avctx, AV_LOG_DEBUG, "Sync to pic %"PRId64"/%"PRId64" "
-           "(recon surface %#x).\n", pic->display_order,
-           pic->encode_order, pic->recon_surface);
+           "(input surface %#x).\n", pic->display_order,
+           pic->encode_order, pic->input_surface);
 
-    vas = vaSyncSurface(ctx->hwctx->display, pic->recon_surface);
+    vas = vaSyncSurface(ctx->hwctx->display, pic->input_surface);
     if (vas != VA_STATUS_SUCCESS) {
         av_log(avctx, AV_LOG_ERROR, "Failed to sync to picture completion: "
                "%d (%s).\n", vas, vaErrorStr(vas));
@@ -237,7 +237,8 @@ static int vaapi_encode_issue(AVCodecContext *avctx,
     }
 
     if (pic->type == PICTURE_TYPE_IDR) {
-        if (ctx->codec->write_sequence_header) {
+        if (ctx->va_packed_headers & VA_ENC_PACKED_HEADER_SEQUENCE &&
+            ctx->codec->write_sequence_header) {
             bit_len = 8 * sizeof(data);
             err = ctx->codec->write_sequence_header(avctx, data, &bit_len);
             if (err < 0) {
@@ -253,7 +254,8 @@ static int vaapi_encode_issue(AVCodecContext *avctx,
         }
     }
 
-    if (ctx->codec->write_picture_header) {
+    if (ctx->va_packed_headers & VA_ENC_PACKED_HEADER_PICTURE &&
+        ctx->codec->write_picture_header) {
         bit_len = 8 * sizeof(data);
         err = ctx->codec->write_picture_header(avctx, pic, data, &bit_len);
         if (err < 0) {
@@ -289,7 +291,8 @@ static int vaapi_encode_issue(AVCodecContext *avctx,
         }
     }
 
-    if (ctx->codec->write_extra_header) {
+    if (ctx->va_packed_headers & VA_ENC_PACKED_HEADER_MISC &&
+        ctx->codec->write_extra_header) {
         for (i = 0;; i++) {
             int type;
             bit_len = 8 * sizeof(data);
@@ -317,6 +320,7 @@ static int vaapi_encode_issue(AVCodecContext *avctx,
             err = AVERROR(ENOMEM);
             goto fail;
         }
+        slice->index = i;
         pic->slices[i] = slice;
 
         if (ctx->codec->slice_params_size > 0) {
@@ -336,7 +340,8 @@ static int vaapi_encode_issue(AVCodecContext *avctx,
             }
         }
 
-        if (ctx->codec->write_slice_header) {
+        if (ctx->va_packed_headers & VA_ENC_PACKED_HEADER_SLICE &&
+            ctx->codec->write_slice_header) {
             bit_len = 8 * sizeof(data);
             err = ctx->codec->write_slice_header(avctx, pic, slice,
                                                  data, &bit_len);
@@ -586,6 +591,10 @@ static int vaapi_encode_step(AVCodecContext *avctx,
     } else if (ctx->issue_mode == ISSUE_MODE_MAXIMISE_THROUGHPUT) {
         int activity;
 
+        // Run through the list of all available pictures repeatedly
+        // and issue the first one found which has all dependencies
+        // available (including previously-issued but not necessarily
+        // completed pictures).
         do {
             activity = 0;
             for (pic = ctx->pic_start; pic; pic = pic->next) {
@@ -601,9 +610,15 @@ static int vaapi_encode_step(AVCodecContext *avctx,
                 if (err < 0)
                     return err;
                 activity = 1;
+                // Start again from the beginning of the list,
+                // because issuing this picture may have satisfied
+                // forward dependencies of earlier ones.
+                break;
             }
         } while(activity);
 
+        // If we had a defined target for this step then it will
+        // always have been issued by now.
         if (target) {
             av_assert0(target->encode_issued && "broken dependencies?");
         }
@@ -631,50 +646,35 @@ static int vaapi_encode_get_next(AVCodecContext *avctx,
         }
     }
 
-    if (ctx->input_order == 0) {
-        // First frame is always an IDR frame.
-        av_assert0(!ctx->pic_start && !ctx->pic_end);
-
-        pic = vaapi_encode_alloc();
-        if (!pic)
-            return AVERROR(ENOMEM);
-
-        pic->type = PICTURE_TYPE_IDR;
-        pic->display_order = 0;
-        pic->encode_order  = 0;
-
-        ctx->pic_start = ctx->pic_end = pic;
-
-        *pic_out = pic;
-        return 0;
-    }
-
     pic = vaapi_encode_alloc();
     if (!pic)
         return AVERROR(ENOMEM);
 
-    if (ctx->p_per_i == 0 || ctx->p_counter == ctx->p_per_i) {
-        if (ctx->i_per_idr == 0 || ctx->i_counter == ctx->i_per_idr) {
-            pic->type = PICTURE_TYPE_IDR;
-            ctx->i_counter = 0;
-        } else {
-            pic->type = PICTURE_TYPE_I;
-            ++ctx->i_counter;
-        }
+    if (ctx->input_order == 0 || ctx->force_idr ||
+        ctx->gop_counter >= avctx->gop_size) {
+        pic->type = PICTURE_TYPE_IDR;
+        ctx->force_idr = 0;
+        ctx->gop_counter = 1;
+        ctx->p_counter = 0;
+    } else if (ctx->p_counter >= ctx->p_per_i) {
+        pic->type = PICTURE_TYPE_I;
+        ++ctx->gop_counter;
         ctx->p_counter = 0;
     } else {
         pic->type = PICTURE_TYPE_P;
         pic->refs[0] = ctx->pic_end;
         pic->nb_refs = 1;
+        ++ctx->gop_counter;
         ++ctx->p_counter;
     }
     start = end = pic;
 
     if (pic->type != PICTURE_TYPE_IDR) {
         // If that was not an IDR frame, add B-frames display-before and
-        // encode-after it.
+        // encode-after it, but not exceeding the GOP size.
 
-        for (i = 0; i < ctx->b_per_p; i++) {
+        for (i = 0; i < ctx->b_per_p &&
+             ctx->gop_counter < avctx->gop_size; i++) {
             pic = vaapi_encode_alloc();
             if (!pic)
                 goto fail;
@@ -688,23 +688,32 @@ static int vaapi_encode_get_next(AVCodecContext *avctx,
             pic->display_order = ctx->input_order + ctx->b_per_p - i - 1;
             pic->encode_order  = pic->display_order + 1;
             start = pic;
+
+            ++ctx->gop_counter;
         }
     }
 
-    for (i = 0, pic = start; pic; i++, pic = pic->next) {
-        pic->display_order = ctx->input_order + i;
-        if (end->type == PICTURE_TYPE_IDR)
-            pic->encode_order = ctx->input_order + i;
-        else if (pic == end)
-            pic->encode_order = ctx->input_order;
-        else
-            pic->encode_order = ctx->input_order + i + 1;
-    }
+    if (ctx->input_order == 0) {
+        pic->display_order = 0;
+        pic->encode_order  = 0;
+
+        ctx->pic_start = ctx->pic_end = pic;
 
-    av_assert0(ctx->pic_end);
-    ctx->pic_end->next = start;
-    ctx->pic_end = end;
+    } else {
+        for (i = 0, pic = start; pic; i++, pic = pic->next) {
+            pic->display_order = ctx->input_order + i;
+            if (end->type == PICTURE_TYPE_IDR)
+                pic->encode_order = ctx->input_order + i;
+            else if (pic == end)
+                pic->encode_order = ctx->input_order;
+            else
+                pic->encode_order = ctx->input_order + i + 1;
+        }
 
+        av_assert0(ctx->pic_end);
+        ctx->pic_end->next = start;
+        ctx->pic_end = end;
+    }
     *pic_out = start;
 
     av_log(avctx, AV_LOG_DEBUG, "Pictures:");
@@ -726,7 +735,7 @@ fail:
     return AVERROR(ENOMEM);
 }
 
-static int vaapi_encode_mangle_end(AVCodecContext *avctx)
+static int vaapi_encode_truncate_gop(AVCodecContext *avctx)
 {
     VAAPIEncodeContext *ctx = avctx->priv_data;
     VAAPIEncodePicture *pic, *last_pic, *next;
@@ -776,7 +785,7 @@ static int vaapi_encode_mangle_end(AVCodecContext *avctx)
         // mangle anything.
     }
 
-    av_log(avctx, AV_LOG_DEBUG, "Pictures at end of stream:");
+    av_log(avctx, AV_LOG_DEBUG, "Pictures ending truncated GOP:");
     for (pic = ctx->pic_start; pic; pic = pic->next) {
         av_log(avctx, AV_LOG_DEBUG, " %s (%"PRId64"/%"PRId64")",
                picture_type_name[pic->type],
@@ -830,6 +839,13 @@ int ff_vaapi_encode2(AVCodecContext *avctx, AVPacket *pkt,
         av_log(avctx, AV_LOG_DEBUG, "Encode frame: %ux%u (%"PRId64").\n",
                input_image->width, input_image->height, input_image->pts);
 
+        if (input_image->pict_type == AV_PICTURE_TYPE_I) {
+            err = vaapi_encode_truncate_gop(avctx);
+            if (err < 0)
+                goto fail;
+            ctx->force_idr = 1;
+        }
+
         err = vaapi_encode_get_next(avctx, &pic);
         if (err) {
             av_log(avctx, AV_LOG_ERROR, "Input setup failed: %d.\n", err);
@@ -858,7 +874,7 @@ int ff_vaapi_encode2(AVCodecContext *avctx, AVPacket *pkt,
 
     } else {
         if (!ctx->end_of_stream) {
-            err = vaapi_encode_mangle_end(avctx);
+            err = vaapi_encode_truncate_gop(avctx);
             if (err < 0)
                 goto fail;
             ctx->end_of_stream = 1;
@@ -930,9 +946,10 @@ static av_cold int vaapi_encode_config_attributes(AVCodecContext *avctx)
     VAProfile    *profiles    = NULL;
     VAEntrypoint *entrypoints = NULL;
     VAConfigAttrib attr[] = {
-        { VAConfigAttribRTFormat        },
-        { VAConfigAttribRateControl     },
-        { VAConfigAttribEncMaxRefFrames },
+        { VAConfigAttribRTFormat         },
+        { VAConfigAttribRateControl      },
+        { VAConfigAttribEncMaxRefFrames  },
+        { VAConfigAttribEncPackedHeaders },
     };
 
     n = vaMaxNumProfiles(ctx->hwctx->display);
@@ -1049,6 +1066,23 @@ static av_cold int vaapi_encode_config_attributes(AVCodecContext *avctx)
             }
         }
         break;
+        case VAConfigAttribEncPackedHeaders:
+            if (ctx->va_packed_headers & ~attr[i].value) {
+                // This isn't fatal, but packed headers are always
+                // preferable because they are under our control.
+                // When absent, the driver is generating them and some
+                // features may not work (e.g. VUI or SEI in H.264).
+                av_log(avctx, AV_LOG_WARNING, "Warning: some packed "
+                       "headers are not supported (want %#x, got %#x).\n",
+                       ctx->va_packed_headers, attr[i].value);
+                ctx->va_packed_headers &= attr[i].value;
+            }
+            ctx->config_attributes[ctx->nb_config_attributes++] =
+                (VAConfigAttrib) {
+                .type  = VAConfigAttribEncPackedHeaders,
+                .value = ctx->va_packed_headers,
+            };
+            break;
         default:
             av_assert0(0 && "Unexpected config attribute.");
         }
@@ -1233,8 +1267,9 @@ static av_cold int vaapi_encode_create_recon_frames(AVCodecContext *avctx)
     ctx->recon_frames->sw_format = recon_format;
     ctx->recon_frames->width     = ctx->surface_width;
     ctx->recon_frames->height    = ctx->surface_height;
-    ctx->recon_frames->initial_pool_size =
-        avctx->max_b_frames + 3;
+    // At most three IDR/I/P frames and two runs of B frames can be in
+    // flight at any one time.
+    ctx->recon_frames->initial_pool_size = 3 + 2 * avctx->max_b_frames;
 
     err = av_hwframe_ctx_init(ctx->recon_frames_ref);
     if (err < 0) {
@@ -1347,6 +1382,11 @@ av_cold int ff_vaapi_encode_init(AVCodecContext *avctx)
     ctx->decode_delay = 1;
     ctx->output_order = - ctx->output_delay - 1;
 
+    // Currently we never generate I frames, only IDR.
+    ctx->p_per_i = ((avctx->gop_size + avctx->max_b_frames) /
+                    (avctx->max_b_frames + 1));
+    ctx->b_per_p = avctx->max_b_frames;
+
     if (ctx->codec->sequence_params_size > 0) {
         ctx->codec_sequence_params =
             av_mallocz(ctx->codec->sequence_params_size);
@@ -1373,16 +1413,32 @@ av_cold int ff_vaapi_encode_init(AVCodecContext *avctx)
         }
     }
 
-    // All I are IDR for now.
-    ctx->i_per_idr = 0;
-    ctx->p_per_i = ((avctx->gop_size + avctx->max_b_frames) /
-                    (avctx->max_b_frames + 1));
-    ctx->b_per_p = avctx->max_b_frames;
-
     // This should be configurable somehow.  (Needs testing on a machine
     // where it actually overlaps properly, though.)
     ctx->issue_mode = ISSUE_MODE_MAXIMISE_THROUGHPUT;
 
+    if (ctx->va_packed_headers & VA_ENC_PACKED_HEADER_SEQUENCE &&
+        ctx->codec->write_sequence_header) {
+        char data[MAX_PARAM_BUFFER_SIZE];
+        size_t bit_len = 8 * sizeof(data);
+
+        err = ctx->codec->write_sequence_header(avctx, data, &bit_len);
+        if (err < 0) {
+            av_log(avctx, AV_LOG_ERROR, "Failed to write sequence header "
+                   "for extradata: %d.\n", err);
+            goto fail;
+        } else {
+            avctx->extradata_size = (bit_len + 7) / 8;
+            avctx->extradata = av_mallocz(avctx->extradata_size +
+                                          AV_INPUT_BUFFER_PADDING_SIZE);
+            if (!avctx->extradata) {
+                err = AVERROR(ENOMEM);
+                goto fail;
+            }
+            memcpy(avctx->extradata, data, avctx->extradata_size);
+        }
+    }
+
     return 0;
 
 fail: